Use of Mendelian randomisation to assess potential benefit of clinical intervention

Author's reply: "Mendelian randomisation" rather than "Mendelian attribution"

14 November 2012

In his rapid response [1], the author makes two points: firstly relating to whether the estimate from Mendelian randomisation can be described as causal, and secondly whether the allocation of genetic variants is random.

With regard to the second point, as the author has stated, genetic variants are not allocated randomly, rather they are inherited from our parents. The concept of Mendelian randomisation requires that the distribution of the variants in the population can be thought of as random with respect to environmental and social factors which may be important confounders. Here the analogy between randomised controlled trials and Mendelian randomisation is not perfect, although the important aspect of the analogy is not the possibility of an individual being allocated to any of the genetic subgroups, but that the quasi-randomised genetic subgroups of the population only differ with respect to the risk factor of interest.

The necessary assumptions for a variant to be randomly distributed are random mating and lack of selection effects relating to the variant(s) of interest. While there will be some departures from these assumptions, studies have shown that the distribution of most genetic variants is fairly uniform across the population despite differences in environmental and social factors, at least in a Western European context [2]. Considerable departures from the assumptions which may invalidate the use of a genetic variant can be assessed by performing a test of Hardy--Weinberg equilibrium, to see if the frequency of heterozygotes and homozygotes in the population is in line with what is expected.

With regard to the first point, Mendelian randomisation is a technique for exploiting genetic variation as a natural experiment [3]. In any natural experiment, the interpretation of any finding as a causal effect requires assumptions of consistency and ignorability [4]. Put simply, these assumptions state that the effect on the outcome of the observed difference in the risk factor of interest due to the genetic variant would be the same if the genetic variant were manipulated to take different values, rather than being observed at different values. Our point in this paper is that in practice this is moot, as the true causal effect of interest is the result of manipulating not the genetic variant, but the risk factor directly. Hence although Mendelian randomisation is an observational rather than an experimental technique, it does estimate a causal parameter (subject to the above assumptions), but this parameter is not necessarily equal to the causal estimand of interest.

As to the name of the technique, "Mendelian randomisation" is justified by the requirement of the genetic variants to be randomly distributed in the population with respect to potential confounders, and by the analogy (albeit not perfect, as discussed here and in the original article) with randomised controlled trials.

1. Vaucher P. "Mendelian attribution" rather than "Mendelian randomisation", BMJ 2012,

2. Davey Smith, G. et al. Clustered environments and randomized genes: a fundamental distinction between conventional and genetic epidemiology. PLoS Medicine, 2007, 4, e352.

3. Glymour MM. Natural experiments and instrumental variable analyses in social epidemiology. Chapter in "Methods in social epidemiology", Jossey-Bass 2006.

4. HernĂ¡n M, Robins J. Causal inference. Chapman & Hall/CRC, 2013.

Competing interests: I co-wrote the original article.

Stephen Burgess, Statistician

Department of Public Health and Primary Care, University of Cambridge, Strangeways Research Laboratory, 2 Worts Causeway, Cambridge, CB1 8RN

Click to like: